Panel Data Break the Diagnostics of Geographically Weighted Regression: Five Failures, Their Repair, and a Rice-Productivity Application in Indonesia

Geographically weighted panel regression (GWPR) is assembled from cross-sectional tools. Five of them fail on a balanced panel whose units keep their coordinates over time. Each failure is stated as a proposition, derived in closed form, and repaired. An adaptive bandwidth in rows is not identified. A fold that holds out a whole unit cannot score a fixed-effects model without a plug-in level. A grid counted in units saturates at the inner training size, where its tie rule picks the bandwidth. A local and a global model share one information criterion only through a composite smoother. The BreuschPagan statistic reduces to a constant set by the sample size. The repairs are applied to 272 province-years, comprising 34 Indonesian provinces from 2018 to 2025, with log yield as the outcome. Under nested leave-one-province-out cross-validation, the global two-way fixed-effects model attains a cross-validated R2 of 0.100 and the GWPR 0.064. The GWPR loses in every variant tried. A province-block permutation test cannot distinguish Indonesia's real map from a random relabelling, and none of the 136 local slopes differs from its global counterpart. The R2 of 0.9997 in the earlier version of this study comes from the accounting identity between production and harvested area.